Knowledge-constrained projection of high-dimensional data

Omanović, Amra (author), Oblak, Polona (mentor), Zupan, Blaž (co-mentor)

Projection of high-dimensional data is usually done by reducing dimensionality of the data and transforming the data to the latent space. We created synthetic data to simulate real gene-expression datasets and we tested methods on both synthetic and real data. With this work we address the visualization of our data through implementation of regularized singular value decomposition (SVD) for biclustering using L0-norm and L1-norm. Additional knowledge is introduced to the model through regularization with the two prior adjacency matrices. We show that L0-norm SVD and L1-norm SVD give better results than standard SVD.

• Language: English
• Keywords: data projection, latent spaces, regularization, data science, single-cell genomics
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2018
• PID: 20.500.12556/RUL-103094
• Publication date in RUL: 13.09.2018

Secondary language

• Language: Slovenian
• Title: Projekcija visokodimenzionalnih podatkov ob upoštevanju domenskih omejitev

Abstract

Projekcija visokodimenzionalnih podatkov se običajno pripravi z zmanjšanjem dimenzionalnosti, ki se predstavi v latentnem prostoru, kar omogoča smiselno vizualizacijo. Pripravili smo sintetične podatke, ki odražajo gensko izražanje v pravih podatkovnih zbirkah. Metode smo kasneje testirali na pripravljenih sintetičnih in pravih podatkih. V tem delu smo obravnavali naloge z izvajanjem regularizirane SVD metode, z uporabo L0-norme in L1-norme. Modelu je bila dodana informacija z regularizacijo dveh dodatnih matrik sosednosti. Pokazali smo, da so te metode dale boljše rezultate kot standardni SVD.

• Keywords: projekcija podatkov, latentni prostori, regularizacija, podatkovna veda, genomika posameznih celic